Even after several years of widespread use of language models, certain misconceptions persist and lead to costly technical decisions.
1. "Model for Retrieving Information Learned"
It doesn’t store documents and then retrieve them. Knowledge is scattered across weights in the form of statistics, so specific details—numbers, dates, and citations—are the most prone to distortion. That’s why RAG exists.
2. "Bigger is always better"
This is generally true for models within the same family. However, a small model that is fine-tuned for a specific task can outperform a large, general-purpose model, at a fraction of the cost and with significantly lower latency.
3. "Tell him not to make things up, and he won't."
The model cannot distinguish when it is making things up—true and false statements are generated by the same mechanism. An effective approach is to provide context for the source and flag it as a quote, rather than offering advice.
4. "At 0 degrees, the result is fixed."
Lowering the temperature makes the output much more stable, but the order of floating-point accumulation on the GPU and the way requests are batching can still cause minor variations. Do not design the system based on the assumption of bit-for-bit reproducibility.
5. "A large context window means making good use of the entire window"
The ability to extract information is best in the middle of a long block of text, rather than at the beginning or end. The order of the documents in the prompt really does make a difference.
6. "Refinement is a way to teach new knowledge"
Fine-tuning teaching methods, formats, and behaviors is very effective. Cramming new knowledge through fine-tuning is both costly and likely to undermine overall competence. Knowledge should be conveyed through context.
7. "Evaluate by reading a few answers"
Based on a subjective assessment of over ten examples, no 5% decline in quality was detected—but a 5% drop on a large scale would mean a significant number of users encountering errors. A fixed test suite is needed that runs every time the prompt or model is changed.
Thảo luận